The goal of this talk is to expose the issues related to getting NERSC MPP's (or, in general, distributed memory machines) using MPI to render 1 billion triangles per second. There are several reasons for wanting to rendering 1 billion triangles per second. One reason is to see if it can be done with the machines at hand and to learn about them and rendering in the process. Another reason is to generate visualizations as large scientific computations are being done using the existing data decomposition. Specifically, within the framework of CACTUS.
To this end, I have been characterizing the NERSC machines and their performance for various operations that need to be done - e.g. coordinate transformations, triangle rasterizing, image/z-buffer composition. I will be talking about my strategies for this, my current results, and other possibilities. The point here is to get feedback from other people at NERSC about my ideas and to get suggestions for better ways to do portions of the overall rendering task.
Snacks will be provided.
See Conundrum Talks for more information about this series.